Sampreeth Avvari
Resume
Projects

Guided tour

NPC Coach. Every new patient call graded against the playbook.

Every new-patient phone call gets transcribed and graded against the practice's documented playbook, producing a coaching report that quotes the exact moment a coordinator missed a non negotiable, next to their own 30 day average.

Speech + LLM pipeline · Call coaching 6 steps · synthetic data

Walk through it, six steps

Step 1 of 6

Two ingestion paths, one call record

Outbound calls arrive as MP3 drops on SFTP; inbound calls arrive from the call platform with a transcript already attached, which the pipeline reuses instead of paying to transcribe twice. Both normalize into one record, claim an idempotency key, and create a visible run row before any filtering, so no call can silently disappear.

IngestSynthetic data
$ ingest SYN-4471.mp3 (inbound · transcript attached)
idempotency key claimed · run row created (visible pre-filter)
classify: new-patient call ✓ · not voicemail · not wrong-number
transcript reused from platform → $0 transcription cost
Step 1 of 6

Every name, transcript and document in this tour is synthetic, and so are the numbers on the sample screens. The numbers under The problem and Where it landed are real. The production system and its data stay private.

The problem.

New-patient coordinators are the practice's first advocates, and their calls set the tone for everything after. But nobody could listen to every call, so coaching was anecdotal. The design constraint came from the top: correctness beats conversion, and booking rate is explicitly not the success metric.

  • Calls graded 690 real calls since June 17
  • Call end to grade 14 min median, October
  • Hard gates 3 of six criteria, a miss raises a red flag
  • Built in 16 days including the React dashboard

How it fits together.

4 layers, left to right

  1. Ingest

    • SFTP MP3 drops + call-platform webhooks
    • One normalized call record, idempotency claims
    • Run rows created before any gate: nothing disappears

  2. Understand

    • Speech-to-Text v2 with diarization (outbound)
    • Platform transcripts reused (inbound), cost deduped
    • Call classification: new patient vs everything else

  3. Score

    • Gemini structured output with response schema
    • 6 criteria, 3 hard gates, weighted buckets 40/40/20
    • Verbatim evidence quotes with a relevance gate

  4. Deliver

    • React dashboard: runs, steppers, logs, analytics
    • Email reports to coordinator + manager
    • Firestore state · BigQuery audit · Terraform · ports-and-adapters demo mode

Where it landed.

690 new patient calls since June 17 have been graded on the same rubric, with receipts. The ports-and-adapters design means the whole pipeline runs on synthetic data in demo mode, which is exactly what this tour is. No real transcripts, recordings or caller details appear here, and coordinators appear as C-1, C-2 and so on.

Concepts

Speech-to-text + diarizationStructured LLM scoringHard gates vs weighted scoresIdempotent ingestionPorts and adaptersEvidence-quoted evals

Stack

PythonFastAPICloud RunSpeech-to-Text v2Vertex AI GeminiFirestoreBigQueryReactTerraform
Read the case study All demos